Hindawi

Mobile Information Systems

Volume 2022, Article ID 5168898, 9 pages
https://doi.org/10.1155/2022/5168898

Research Article
A Method of Key Posture Detection and Motion Recognition in
Sports Based on Deep Learning

Shaohong Pan
Xichang University, Xichang City, Sichuan 615000, China

Correspondence should be addressed to Shaohong Pan; xcc03300051@xcc.edu.cn

Received 16 February 2022; Revised 20 March 2022; Accepted 21 March 2022; Published 25 April 2022

Academic Editor: Tongguang Ni

Copyright © 2022 Shaohong Pan. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Moving target recognition and analysis is an important research direction in the field of computer vision, which is widely used in
our life, such as intelligent robot, video surveillance, medical education, sports competition, and national defense security. By
analyzing the video of weightlifting, this paper extracts the key postures of athletes’ training, so as to assist coaches to train
athletes more professionally. Based on DL (Deep Learning), a key pose extraction method of sports video (RoI_KP for short)
based on classified learning of regions of interest is proposed. By fine-tuning CNN (Convolutional Neural Network), a network
model suitable for video classification of weightlifting in the region of interest is obtained. Finally, according to the
classification results, the selection strategy of classification results is designed to extract key poses. According to the
characteristics of different modal information, different DNN (deep neural network) is adopted, and various depth networks
are combined to mine the multi-modal spatio-temporal depth features of human movements in video. Experimental results
show that the method proposed in this paper is very competitive.

1. Introduction detect abnormal behavior in wards and notify the hospital

Video analysis technology is the basis of content-based video
retrieval. In the traditional text-based query technology,
query keywords can basically reflect the query intention
[1]. However, in content-based video retrieval, there is a dif-
ference between the underlying features and the upper
understanding, mainly because the underlying features can-
not fully reflect or match the query intention [2]. At the
same time, human movements in sports videos are very
complicated and skillful, and compared with daily sports
[3], the analysis of sports videos is more difficult and chal-
lenging [4]. Therefore, the analysis of sports videos can not
only bring more viewing effects to sports competitions, but
also analyze competitions between athletes and coaches
and assist athletes in training. In comparison to manual
monitoring, the computer does not need to rest, and it will
not miss important information monitoring due to a variety
of external factors, allowing it to significantly reduce man-
power and material resources while improving work effi-
ciency. This technology is also essential in the medical
field. Ward monitoring, for example, can automatically
in real time, which not only reduces the risk of a patient’s
sudden situation, but also lowers the cost of medical treat-
ment. The query condition in content-based video retrieval
is an image sequence or a description of video content. To
create an index, first extract the underlying features, then
calculate and compare the distance between these features
and the query condition to see if they are similar.
Intelligent license plate recognition, real-time 3D effect
playback of sports competitions, and other technologies
have emerged one after another, which makes the advan-
tages of computer vision task based on machine learning
[5–8] more and more obvious, and it has high flexibility
and openness, which indicates the development direction
of intelligent image processing [9–11]. In the process of
weightlifting, whether the athletes’ movements are standard
or not will directly affect the athletes’ performance. In the
whole process of weightlifting, there are several key postures
that will affect the athletes’ performance. These key postures
are very important for the success of weightlifting. However,
due to the limitation of the hardware conditions of the mon-
itoring equipment (camera) and the influence of
2 Mobile Information Systems
environmental factors, there are many complicated and
unavoidable interference factors when acquiring surveillance
video, such as various occlusion between objects, noise,
background influence, and sudden change of light, which
makes it very challenging for us to use target detection and
tracking in intelligent monitoring and other related applica-
tions. However, the recognition of motion and weight in
video is also facing a huge test: the number of super-large-
scale videos requires higher and higher computing perfor-
mance of the algorithm. At the same time, with the emer-
gence of a large number of video content such as aerial
photography and first-person perspective, the shortcomings
of traditional algorithms in dealing with different camera
perspectives, cluttered background, occlusion, and other
issues are becoming more and more obvious.
In the training process of athletes, the judgment of key
postures in previous weightlifting videos was mainly based
on coaches’ experience, which not only wasted manpower,
but also the extracted key postures were not accurate and
easily influenced by subjective factors. Facing weightlifting,
this paper applies video analysis technology to weightlifting
training, and analyzes the movement of weightlifting pro-
cess, aiming at putting forward a reliable automatic extrac-
tion method of key posture frames in the movement
process, which is of far-reaching significance for athletes
and coaches to master movement techniques as soon as pos-
sible, improve training efficiency, and improve training level.
2. Related Work
Literature [12] detects the edge of the input video stream,
uses the edge features to establish gesture descriptors, and
then clusters the features calculated by the video stream to
get the important gestures. In this way, good model param-
eters can be learned through loop iterative modeling. Litera-
tures [13, 14] apply SVM (Support Vector Machine) to
human motion recognition. They use local spatio-temporal
features to represent human motion, and then input local
spatio-temporal feature vectors into SVM to judge the cate-
gory of human motion. Literature [15] extracts spatial fea-
tures from static points of interest in video, and extracts
spatio-temporal features from dynamic points of interest at
the same time, and obtains a composite feature set contain-
ing both static spatial features and dynamic spatio-temporal
features, which is input into linear SVM for classification. A
proposed algorithm in the literature [16] is as follows: The
gesture features were obtained by training with the edge fea-
tures obtained after the video was edge detected. The motion
was then approved by a majority of the members. Literature
[17] proposes a full-body and half-body model, in which the
body model employs densely sampled shape context descrip-
tors and the prior model is exhaustively trained in the data-
base using multi-angle and multi-pose. Finally, the image’s
features are used to detect human body parts in the image,
which is determined by the gesture space representation
and inference algorithm.
Literature [18] introduced the spatial pyramid pool layer
into CNN (Convolutional Neural Network) and proposed
SPPnet, which reduced the limitation of CNN network on
the size of input pictures and improved the accuracy. On this
basis, literature [19] puts forward a faster detection method
of Faster-RCNN, and puts forward a new innovation, that
is, extracting the RPN (Region Proposal Network) of candi-
date regions, thus completing the technical evolution from
RCNN to Faster-RCNN. Literature [20] puts forward a
new detection method, and its performance is greatly
improved compared with the traditional one-stage and
two-stage framework, especially in the real-time situation,
and its accuracy is much higher than YOLOV3 at the same
rate. Literatures [21, 22] use NN (nearest neighbor) classifier
to judge whether tracking is successful or not. Neural net-
work classifier can measure the similarity between the col-
lected correct image and the current new target image. The
DNN (deep neural network) and detector in the tracking-
learning-detection framework were improved in literature
[23]. The tracking failure point is identified by calculating
the DNN’s forward-backward error and the output’s
spatial-temporal similarity. Literature [24] uses the twin net-
work to directly calculate the representation error of the tar-
get appearance model, with the goal of selecting the region
that is closest to the target model as the tracking result.
HRNet is an end-to-end multi-person 2D pose estimation
method proposed in the literature [25]. The top-down
method has some flaws, such as misalignment of key points
in crowded situations or inaccurate detection in bust situa-
tions, based on its characteristics. According to the literature
[26], organizing two-dimensional posture nodes by guiding
all pelvic joints can be done quickly and with good results.
If the person in the video is performing fast and complex
actions, the estimated two-dimensional posture will be
unstable, resulting in motion distortion.
3. Research Method
3.1. Sports Key Posture Detection. In order to detect the tar-
get from the image in any scene, the key is to choose a robust
feature set, and apply this feature set, so that the targets of
different classes have good discrimination and can adapt to
the differences of the same class of targets. The features used
for target recognition mainly include shape, texture, and
color, while the shape of the target is not affected by factors
such as illumination, and it is often used as a good feature
for target recognition. Different descriptions of the shapes
of objects produce different feature sets.
The process of weightlifting is generally divided into five
stages: extending the knee to lift the bell, leading the knee to
lift the bell, exerting strength, squatting and supporting, and
standing up. Coaches and athletes want to observe the
movement process of each movement, and hope to observe
biomechanical parameters such as the movement track of
each snatch or clean and jerk, the speed of each stage, and
the performance of work. It is impossible to use traditional
pose estimation methods to precisely locate the positions
of various parts of the human body, and then extract the
motion key frames according to the attitude estimation
method. Moreover, there is a serious problem—the similar-
ity between frames. Because of the continuity of motion,
the difference between adjacent video frames is very small.
Mobile Information Systems
In this chapter, a key pose extraction method of sports
video (RoI_KP for short) based on region of interest classifi-
cation learning is proposed. The algorithm flow is shown in
Figure 1.
First, frame the video. Then, based on the first frame of
the video, the trained model is used to segment and extract
the foreground, and the video is segmented according to
the standard of the first frame. The region of interest directly
related to key frame extraction is obtained. Furthermore,
CNN network is used to extract and classify the features of
each video frame to obtain candidate key frames. Finally,
the key frame extraction strategy is formulated, and key
frames are selected according to the probability value of
the corresponding class output by each frame.
Because there is a lot of background interference infor-
mation in weightlifting video, this paper extracts the region
of interest to reduce the influence of background on key
frame extraction. Then, divide and fine-tune the region: tra-
verse the image with four corners of the image as the starting
points in turn, and update the label value of each pixel point
to the minimum value of the label values of four neighboring
points that is not zero until the label values of all points do
not change;

 joint points from the previous moment.
labelx,y = min label
label ∈ LABELx,y : ð1Þ
Up to this point, a new annotation map of the region of
interest has been obtained, and each non-zero and continu-
ous region has a different number. Finally, only the region
with the largest number needs to be selected as the region
of interest.
Unlike 2D human pose estimation, 3D human pose esti-
mation based on DL is more challenging. This is mainly
because 2D pose estimation has a wider training data set,
so it can better solve occlusion and accuracy, while 3D pose
estimation has great challenges in occlusion and accuracy
due to the lack of training data set.
Three-dimensional human posture estimation is to esti-
mate the three-dimensional coordinates ðx, y, zÞ of the related
nodes from pictures or videos, which is essentially a regression
problem. It is widely used in motion capture system, anima-
tion, behavior understanding, and games. It can also be used
as an auxiliary link of other algorithms such as pedestrian rec-
ognition, and it can also be combined with other tasks related
to human body such as human body analysis.
Each block is based on a simple, deep, and multi-layer
neural network with batch normalization function. After
obtaining 2D key points, input them into 3DPoseNet and
output the estimated 3D coordinate key points, each of
which is expressed as Pi = jX i jjY i jjZ i j. Assuming that 3D
joint ground truth is available, 3D joint MSE (Mean Square
Error) loss is adopted as the loss function by 3D 3DPoseNet:
 2
LMES = 〠Pi − Pi′2 : ð2Þ
Among them, Pi′is the marked true coordinate of the 3D
skeleton, while Pi is the corresponding predicted coordinate,
3
and the coordinates of key points of the 3D skeleton pre-
dicted by 3DPoseNet.
In this paper, the loss function with symmetric con-
straint is used, and the loss function defined in this paper
is formula (3).

   2 
Lsym = 〠 Pi − P j 2 − P
 i′ − P ′
j , ð3Þ
2 2
ði, jÞ∈E
where E is the set of all adjacent points, and Pi′, P′j represents
the key point of the symmetric part. Of course, there are
other restrictions on human bones, such as the limitation
of joint angle.
This paper presents a general solution to the occlusion
problem that makes extensive use of the torso, limbs, and
head. Because 2D posture can accurately predict 3D posture,
the restored 3D posture can be divided into four states,
including whether the limbs are covered and whether the
posture of the entire limbs can be predicted. This paper
introduces the developing concept of time series information
[17] in this case. The difference is that this paper only needs
to determine some position data, i.e., the unpredictable
(occluded part) joint position can be determined using the
Pti = λPt−1
i : ð4Þ
In which, the position t of the joint point i (occlusion
node) of the predicted Pti at the time point is called
ðjX i jjY i jjZ i jÞT , and Pt−1
i represents the position at t − 1 time.
λ is a vector, which represents the general goal of the move-
ment direction of people as objects.
3.2. Motion Recognition Method Based on DL. With the
rapid development of computer hardware and the successful
application of DL in various fields of computer vision, moving
target recognition based on DL has also developed into a key
technology in the field of computer vision, which is widely
used in medical treatment, transportation, security, and so
on, and can also be used as the basis of other application tech-
nologies, such as image processing and three-dimensional
modeling. The traditional methods of target detection can be
divided into three steps: selecting candidate areas based on
images, extracting visual features, and classifying by a class
of commonly used classifiers such as SVM model. With the
development and application of DL, it not only simplifies the
complexity of traditional methods, but also improves the
detection performance. This kind of target detection methods
is generally based on region extraction, which can be divided
into two types: one-stage method and two-stage method.
Obtaining 3D coordinates from video is extremely diffi-
cult. Another method is to represent human motion directly
using 2D human posture changes. Poses of the two-
dimensional human body in different viewpoints have differ-
ent forms. Viewpoint normalization, multi-viewpoint tra-
versal, and viewpoint invariant feature extraction are three
common techniques for achieving viewpoint-independent
representation of human motion. The camera’s position
4 Mobile Information Systems
Video
First
frame?
Region of interest
extraction
Feature extraction
Candidate attitude
recognition
Key frame extraction
strategy
Figure 1: Algorithm framework.
should be fixed, but this is not always the case in sports
videos, because the target moves very quickly, and the cam-
era often moves with it to keep up with the target’s move-
ment, or the moving target will disappear. When filming a
diving video, for example, the camera will move down as
the human body falls. This chapter proposes a multi-modal
information-based method for action recognition. Figure 2
depicts the general framework of this method.
The static information used in this paper is RGB map
and depth map. RGB map contains global information of
environment and human body, which can be used to elimi-
nate background clutter and other problems. On the basis of
the learning model based on DNN, different network struc-
tures are designed for different modes, so as to discover the
temporal and spatial depth characteristics of motion video
and improve the performance of motion recognition model.
In this chapter, CNN network with the same structure is
used for RGB map and depth map. CNN for RGB map rep-
resents the color and texture features of human body and
background in the video, while CNN for depth map accu-
rately distinguishes the front and back scenes of the video
to prevent the misunderstanding of features caused by back-
ground interference.
At the top of the network, this chapter uses Softmax loss
function. Assuming that the output of the last full connec-
tion layer is a vector xi of K dimension, the Softmax function
is defined as:
1 tively. Compared with single modal information, the
σðxi Þ = exp ðxi Þ : ð5Þ
∑Ki=1 expðxi Þ
Y Automatic segmentation
of region of interest
Video analysis
Independent RGB map features and depth map features
can be obtained. On the basis of this model, the output of the
first fully connected layer can be regarded as a convolution
feature with good representation effect, and it has classifica-
tion and discrimination characteristics.
Human motion in video has time characteristics, so only
mining spatial depth features cannot express the temporal
characteristics of motion. The networks that deal with tim-
ing information in common DL methods are all based on
the architecture of RNN (recurrent neural network). In the
task of human motion recognition, the trajectory of key skel-
eton points has effective time information.
RNN network processes sequence data in an iterative
way according to the time scale of the sequence. This pro-
cessing method makes RNN have obvious advantages in
modeling and feature extraction of sequence data. This
method uses the cross entropy loss function:

E yt , yt′ = −yt log yt′ = −〠 yt log yt′ ð6Þ
t
where yt is the correct label at time t, and yt′is the network
prediction label. The training goal is to make the value of
the loss function reach a lower level by calculating the gradi-
ent optimization parameter W of the loss function.
The static model and the dynamic model learn the spa-
tial and temporal characteristics of the action video, respec-
features under multi-modal information have more differ-
ences and complementarities, so the fusion of spatial and
Mobile Information Systems
RNN
unit
Action sequence
RNN
unit
R
CNN
G B
RGB map
CNN
Depth map
Figure 2: Architecture diagram of deep space-time feature motion recognition.
temporal features can provide information with stronger
representation ability. The goal of this experiment is to
directly concatenate temporal and spatial depth features
and estimate the weights of the two features in a linear com-
bination process based on the test accuracy of the two net-
work models, in order to identify the importance of
different features.
In the late stage of fusion, the probability outputs using
spatial and temporal depth features are superimposed by lin-
ear weighting to get the final predicted value. Specifically, the
late fusion adopts the following methods:
! backgrounds or even the same actions. At the same time, it
N
1
P= 〠 αP1 + ð1 − αÞP2
N i=1
α is the weighting parameter, P1 , P2 is the output proba-
bility of the network using spatial depth feature and tempo-
ral depth feature, respectively, and P is the final prediction
probability, where N is the number of sample features of
the video after multiple sampling.
4. Results Analysis and Discussion
The action behavior of human body is not only distinguish-
able in image space, but also distinguishable in time series.
The tasks of image recognition and detection need to mine
the spatial features, while the video increases the informa-
tion of time dimension relative to the image. Therefore, a
video human motion recognition algorithm needs to dig
deep into its temporal and spatial features. A segment of
video contains many frames.
If all the video tilts are used for a video, it will be a task
that consumes a lot of computing time. At the same time,
because not all frames are related, the recognition effect will
5
Temporal and spatial characteristics of depth
Depth time
characteristics
Classifier
Depth space
feature
be reduced. Therefore, it is a very important task to find the
most distinctive spatio-temporal features in video, which can
improve the accuracy of the above algorithms.
Testing with 512 image frames, the correct recognition
rate of the four poses is shown in Figure 3.
The experimental results show that the algorithm pre-
sented in this paper achieves a good recognition result for
each category. The incorrectly divided frame image of a knee
joint posture is output along with the probability value for
each category. All of the incorrectly divided results are found
to belong to the first category of knee joint or the third cat-
egory of force posture. Different scenes may have different
ð7Þ should be noted that if the human body’s clothing matches
the video’s background color, problems such as video expo-
sure caused by outdoor lighting and weather will make it dif-
ficult to distinguish actions and behaviors. The same video
may not have a static background, and the task will become
more difficult as the background changes. Figure 4 depicts
the probability change curves of four categories for each
video frame.
The RoI_KP method proposed in this paper has greatly
improved the performance. It can be seen. Traditional
CNN classification algorithm will have great fluctuation in
classification performance, and it will be unstable in
classification.
With the emergence of various new shooting methods,
such as self-timer, aerial photography, and first-angle shoot-
ing, it is impossible for video information to always keep the
same camera angle. Therefore, the same video action
sequence may cause different feature representations due to
different camera angles. Therefore, the identification of dif-
ferent camera angles is also an urgent problem to be solved.
The experimental input is set as a continuous and orderly
video meal, so the representation of features has all the
6 Mobile Information Systems
Number of right and wrong
1
0.8
Probability value
0.6
0.4
0.2
0
0 5
Each frame image
Stretch knees
Knee leading
Figure 4: Test video results.
spatial information, and at the same time contains part of
the temporal information. However, there is little pre-
training information of depth map information, so the net-
work is trained from scratch. Figures 5 and 6 are the com-
parison results of two groups of videos.
While the traditional CNN algorithm shows great fluctu-
ation in the fourth key frame and the probability values of
other poses are relatively large, which has an impact on the
classification results, the traditional CNN algorithm has a
good effect in the fourth key frame and the probability value
of the fourth key pose is very large. The pre-training process
is not used for RNN but used for bone point information
because the processed features obtained after pre-
180 98.00%
160 97.50%
140
97.00%
120
96.50%
Accuracy
100
96.00%
80
95.50%
60
95.00%
40
20 94.50%
0 94.00%
Posture 1 Posture 2 Posture 3 Posture 4
Posture
Correct
Wrong
Accuracy
Figure 3: Test accuracy of four key frames.
processing are sufficient for training. Learning transfer
allows researchers to obtain good training results even when
the new data set is small, labeling is incomplete or small, and
the distribution of the training and test sets of data is differ-
ent. Low-level features can also be used for similar tasks
because the representation of convolution kernels for CNN
is a process from low-level to high-level and from general
to special.
In this paper, the 3DPoseNet network is evaluated
according to the public data set Human3.6M, and the other
two data sets are used for testing. Human3.6M is the most
widely used public data set in 3D pose at present, which con-
tains 3.6 million video frames for 11 topics, among which 7
10 15 20
topics use 3D annotation pose, each object performs 15
actions, and the video frames (video with frame rate of
50fps) are captured by four cameras with different angles
Power of view, as shown in Figure 7.
Culmination It demonstrates that the 3DPoseNet model’s results have
a lower average error than all other methods and are unaf-
fected by other data or operation methods. The average error
of MPJPE in the 3DPoseNet network of this method is
reduced by at least 6 mm, demonstrating the method’s effec-
tiveness. In general, if a human posture is similar at two dif-
ferent moments of an action, the distance between the two
descriptor features describing the human posture will be
small; on the other hand, if the human posture is very differ-
ent at two different moments, the distance between the two
descriptor features will be large, and this property has little
relation with the viewpoint position of observing the move-
ment. The self-similar matrix with the same characteristics
has similar patterns, as shown in Figure 7, indicating that
changes in sports performers have little impact on the self-
similar matrix. The difference in self-similar matrix based
on different features is very obvious from a vertical
Mobile Information Systems 7

1 85

0.8 75
Probability value

0.6
65

Error
0.4
55
0.2
45
0
0 5 10 15 20
35
Each frame image ref (15) ref (17) ref (19) ref (22) ref (23) Our
Stretch knees Power Methods
Knee leading Culmination
Eat Photo
Figure 5: CNN test results. Greet Pose
Phone

1 Figure 7: Results of average position error per joint.

0.8
60
Probability value

0.6 55

50
0.4
45
Error

0.2 40

35
0
0 5 10 15 20
30
Each frame image
25
Stretch knees Power ref (15) ref (17) ref (19) ref (22) ref (23) Our
Knee leading Culmination
Methods
Figure 6: Results based on RoI_KP. Eat Photo
Greet Pose
perspective, so the self-similar matrix is dependent on the Phone
image’s underlying features. As shown in Figure 8, it can
Figure 8: Joint position error based on Procrustes analysis.
be clearly found that the accuracy of this method is greatly
improved.
All the videos of human body movements were shot in based on a large amount of training data. Different depth
fixed scenes. In order to prevent the moving target from los- networks have different characteristics. CNN network pays
ing from the field of vision, moving lens is often used to more attention to the relationship between local informa-
shoot, and it is difficult to detect the movement of human tion, so it is suitable for image recognition and detection
body from the optical flow field. Figure 9 shows the process tasks.
of 3D CNN in pre-training stage and network fine-tuning Experiments show that the self-similar matrix of human
stage, respectively. It can be seen that the fluctuation and posture obtained by using directional gradient histogram of
loss output are larger in the training process, and the net- human foreground image to represent human posture in dif-
work output loss is reduced after fine-tuning, thus improv- ferent viewpoints has better stability. By applying the filter to
ing the final classification effect of the model and reducing the post-processing of 3D pose attitude, the predicted atti-
the time cost of network learning. tude stability can be effectively improved. On the basis of
The action recognition method based on DL in single successful motion capture, through the analysis and extrac-
mode includes CNN and recurrent neural network. It also tion of human motion feature parameters, we can automat-
explains the problems and difficulties brought by single- ically analyze and understand various human motions and
mode video motion recognition, and introduces multi- behaviors. Motion recognition technology has a wide appli-
mode motion recognition method. DL gets better results cation prospect and great economic and social value in
8 Mobile Information Systems
× 104
6
5
Loss of function output
4 References
3
2
1
0 [3] M. Gao, W. Cai, and R. Liu, “AGTH-Net: attention-based
0 50
Iteration
Pre-training
Fine-tune
Figure 9: 3-D CNN pre-training and fine-tuning network loss.
advanced human-computer interaction, rehabilitation engi-
neering, sports analysis, somatosensory game control, and
content-based retrieval.
5. Conclusion
As the most important element in the human environment,
rich and diverse human movements carry a great deal of
information which is extremely important for human social
interaction. Therefore, studying human movements has pro-
found economic and social value. The research of human
motion recognition based on vision involves multi-
disciplinary knowledge, and integrates the related research
results of computer vision, pattern recognition, image pro-
cessing, machine learning, and many other disciplines.
When we regress 2D to 3D pose nodes based on DL, we
can effectively improve the accuracy of the network model
by calculating the scale of 2D to 3D node coordinates of each
frame based on the least square method. The human body in
weightlifting video is extracted by skeleton to enhance the
expression of features, thus further improving the accuracy
of key frame extraction. Deep CNN is used to mine the spa-
tial characteristics of static information, improve the repre-
sentation of dynamic information, and recursive neural
network is used to process it. Experimental results show that
the method proposed in this paper is very competitive.
From 2D pose estimation to 3D pose prediction, if there
is occlusion, it is mainly self-occlusion, which will greatly
affect people’s visual effect. Therefore, the focus of our future
work is to better solve the occlusion problem and improve
the processing speed of the framework.
Data Availability
The data used to support the findings of this study are avail-
able from the corresponding author upon request.
Conflicts of Interest
The author does not have any possible conflicts of interest.
[1] J. Miller, U. Nair, R. Ramachandran, and M. Maskey, “Detec-
tion of transverse cirrus bands in satellite imagery using deep
learning,” Computers & Geosciences, vol. 118, pp. 79–85, 2018.
[2] H. Liu, J. Jin, Z. Xu, Y. Bu, Y. Zou, and L. Zhang, “Deep learn-
ing based code smell detection,” IEEE Transactions on Soft-
ware Engineering, vol. 99, 2021.
100 150 graph convolution-guided third-order hourglass network for
sports video classification,” Engineering, vol. 2021, pp. 1–10,
2021.
[4] Z. Wang, Z. Zhou, H. Lu, and J. Jiang, “Global and local sensi-
tivity guided key salient object re-augmentation for video
saliency detection,” Pattern Recognition, vol. 103, no. 2, article
107275, 2020.
[5] J. Zhou, Y. Wang, and W. Zhang, “Underwater image restora-
tion via information distribution and light scattering prior,”
Computers and Electrical Engineering, vol. 100, article
107908, 2022.
[6] X. Gu, W. Cai, M. Gao, Y. Jiang, X. Ning, and P. Qian,
“Multi-Source Domain Transfer Discriminative Dictionary
Learning Modeling for Electroencephalogram-Based Emo-
tion Recognition,” in IEEE Transactions on Computational
Social Systems, pp. 1–9, Institute of Electrical and Electron-
ics Engineers, 2022.
[7] M. Zhao, A. Jha, Q. Liu et al., “Faster mean-shift: GPU-
accelerated clustering for cosine embedding-based cell seg-
mentation and tracking,” Medical Image Analysis, vol. 71, arti-
cle 102048, 2021.
[8] W. Cai, B. Zhai, Y. Liu, R. Liu, and X. Ning, “Quadratic poly-
nomial guided fuzzy C-means and dual attention mechanism
for medical image segmentation,” Displays, vol. 70, article
102106, 2021.
[9] J. Zhou, X. Wei, J. Shi, W. Chu, and W. Zhang, “Underwater
image enhancement method with light scattering characteris-
tics,” Computers and Electrical Engineering, vol. 100, article
107898, 2022.
[10] D. Yao, Z. Zhi-li, Z. Xiao-feng et al., “Deep hybrid: multi-graph
neural network collaboration for hyperspectral image classifi-
cation,” Defence Technology, vol. 5, 2022.
[11] J. Zhou, D. Liu, X. Xie, and W. Zhang, “Underwater image res-
toration by red channel compensation and underwater median
dark channel prior,” Applied Optics, vol. 61, no. 10, pp. 2915–
2922, 2022.
[12] A. Lamas, S. Tabik, P. Cruz et al., “MonuMAI: dataset, deep
learning pipeline and citizen science based app for monumen-
tal heritage taxonomy and classification,” Neurocomputing,
vol. 420, pp. 266–280, 2021.
[13] S. Nandyal and S. L. Kattimani, “Bird swarm optimization-
based stacked autoencoder deep learning for umpire detection
and classification,” Scalable Computing, vol. 21, no. 2, pp. 173–
188, 2020.
[14] X. Zhao, T. Zuo, and X. Hu, “OFM-SLAM: a visual semantic
SLAM for dynamic indoor environments,” Mathematical
Problems in Engineering, vol. 2021, 16 pages, 2021.
Mobile Information Systems 9

[15] Q. Guo, “Detection of head raising rate of students in class-

room based on head posture recognition,” Traitement du Sig-
nal, vol. 37, no. 5, pp. 823–830, 2020.
[16] P. Wang, “Research on sports training action recognition
based on deep learning,” Scientific Programming, vol. 2021,
Article ID 6878, 8 pages, 2021.
[17] A. Bruno, F. Gugliuzza, R. Pirrone, and E. Ardizzone, “A
multi-scale colour and keypoint density-based approach for
visual saliency detection,” Access, vol. 8, pp. 121330–121343,
2020.
[18] Y. Xu, “A sports training video classification model based on
deep learning,” Scientific Programming, vol. 2021, Article ID
2896, 11 pages, 2021.
[19] J. Liu, D. Chen, Y. Wu, R. Chen, P. Yang, and H. Zhang,
“Image edge recognition of virtual reality scene based on
multi-operator dynamic weight detection,” Access, vol. 8,
pp. 111289–111302, 2020.
[20] F. A. Memon, U. A. Khan, A. Shaikh, A. Alghamdi, P. Kumar,
and M. Alrizq, “Predicting actions in videos and action-based
segmentation using deep learning,” Access, vol. 9, pp. 106918–
106932, 2021.
[21] G. Kalakoti and G. Prabakaran, “Key-frame detection and
video retrieval based on DC coefficient-based cosine orthogo-
nality and multivariate statistical tests,” Traitement du Signal,
vol. 37, no. 5, pp. 773–784, 2020.
[22] H. Gu, J. Zhang, T. Liu et al., “DIAVA: a traffic-based frame-
work for detection of SQL injection attacks and vulnerability
analysis of leaked data,” IEEE Transactions on Reliability,
vol. 69, no. 1, pp. 188–202, 2020.
[23] Z. Kalal, K. Mikolajczyk, and J. Matas, “Tracking-learning-
detection,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 34, no. 7, pp. 1409–1422, 2012.
[24] B. Pu, N. Zhu, K. Li, and S. Li, “Fetal cardiac cycle detection in
multi-resource echocardiograms using hybrid classification
framework,” Future Generation Computer Systems, vol. 115,
no. 3, pp. 825–836, 2021.
[25] J. Cui, M. Wang, Y. Luo, and H. Zhong, “DDoS detection and
defense mechanism based on cognitive-inspired computing in
SDN,” Future Generation Computer Systems, vol. 97, pp. 275–
283, 2019.
[26] C. Li-quan, L. You, F. Shen, Z. Shan, and J. Chen, “Pose recog-
nition in sports scenes based on deep learning skeleton
sequence model,” Journal of Intelligent and Fuzzy Systems,
vol. 3, pp. 1–10, 2021.